This one is a bit later than I'd expected, but hopefully it is worth the wait!

I really enjoyed Algorithms to Live By: The Computer Science of Human Decisions by Brian Christian and Tom Griffiths — it's my favourite so far this year. It was crammed full of (imo) interesting information so I encourage you to read the book if you're a geek like me.

The key takeaways are below, split into two parts. The first part covers the main messages from the book, explaining how general principles from computer science might help us in our daily lives. The second part outlines specific computer science problems discussed in the book. To keep my summary (somewhat) concise, I’ve split some of these out into separate posts.

As usual, you can find the full summary here.

KEY TAKEAWAYS

• We can apply algorithms to some of the problems we face in real life.
• However:
  • Even the best algorithm can have a pretty high failure rate. Having the best process does not guarantee the best outcome.
  • Real-life problems are hard, and many are intractable (i.e. no efficient solution exists).
• Yet there is still hope:
  • Humans do surprisingly well at difficult computer science problems.
  • Sometimes less is more, and simple models can outperform complex ones that overfit the data.
  • Relaxing an intractable problem can make it tractable.
  • Randomness can also help produce “good enough” solutions much quicker.
• Computer science manages trade-offs. For example:
  • Sorting pre-empts the costs of a future search, so there is a trade-off between sorting and searching.
  • Caching manages the trade-off between time and space.
  • Queues bring up average throughput at the expense of delay or latency. When a queue or a to-do list gets too long, it’s better to close it off and drop things entirely.
  • Relaxing a problem may trade off certainty for time/computational resources.
  • In real life, we can try to be computationally kind to others. Constraining someone’s choices can make a decision easier for them.

Specific computer science problems and issues:

• ​Optimal stopping problems. Look at the first 37% of options, then immediately leap at the next best one.
• Explore-exploit trade-offs. An unexplored option has a high value, because it gives us information for future decisions. But this exploration premium declines over time.
• ​Sorting. Sorting gets much harder as the number of items increases. There are massive diseconomies to scale.
• Caching. Removing the least recently used item from your cache is optimal.
• ​Scheduling. Precedence constraints can make problems intractable, while the ability to switch tasks can make it tractable. However, context-switching has a cost. Too much context-switching can lead to thrashing, where nothing gets done.
• ​Bayes’s Rule. We can estimate probabilities by combining pre-existing beliefs (or priors) with new/observed evidence. The prediction rules to apply differ depending on whether something follows a normal, power-law or Erlang distribution.
• Networking. Sometimes we need to back off or drop things to prevent a network or resource from getting overloaded.
• ​Game theory. People acting rationally can lead to worse outcomes for all. We can change those outcomes by changing the rules of the game. The price of anarchy tells us whether a centralised approach would be better than a decentralised one.

Read the full summary for this and more on ToSummarise.com​